Isolated word endpoint detection using time-frequency variance kernels

A major challenge in developing endpoint detection systems is the presence of background noise. We have developed a hybrid method for performing endpoint detection which is based on spectrogram estimation using LPC and a detection process based on imaging operations on the spectrogram. High-variance regions in the spectrogram, captured by variance kernels, can be used to accurately determine the endpoints of speech. This hybrid approach to endpoint detection is robust to various types and levels of background noise. Compared with two other publicly-available methods, our approach performs favorably.